We released a foundation model with everything open — weights, the training data recipe, the full training code, all 162,000 steps of logs. Two days later it showed up in Chinese tech media.
I went through the coverage one article at a time, expecting the usual: a number or two mangled, one impressive figure inflated into a headline.
That is not what I found. What I found was more interesting than the articles themselves.
First: it was remarkably accurate
The original was CNMO (手机中国), a real Chinese tech publication. July 20, 15:32, bylined.
I checked every figure against our model card:
6.59B total parameters · ~2.98B active per token · 49 layers · 7×7 Latin square placement · 16 × NVIDIA B200 · 144.2B training tokens · math 37.8% / Korean 21.6% / English 21.6% · Apache-2.0
Twelve numbers. Not one wrong.
No inflation either. No "world first," no "breakthrough." Our source material didn't contain those words, so the article didn't either.
The sentence the reporter added
One line stayed with me:
并非以英语为中心的模型，而是同时考虑韩语和英语的自主基础模型
"Not an English-centric model, but a sovereign foundation model that considers Korean and English together."
That framing wasn't in any press material. The reporter looked at "Korean 21.6%" and drew the conclusion themselves.
A Chinese outlet reaching for 自主 — sovereign, self-determined — to describe a Korean model is worth pausing on. It is the same word Chinese media use for their own domestic AI.
Someone in the same position recognised the position.
Then I noticed the byline problem
All six articles call the company Bidraft.
The name is VIDRAFT. With a V.
The Korean company name reads bi-deu-raef-teu, and romanising that first syllable back out of Korean gives you B. It's a natural mistake for anyone reading Korean. It's also a costly one.
A Chinese developer who finds the article interesting and searches "Bidraft" gets nothing. Not the homepage, not the Hugging Face org, not a single other article.
Six placements, and no address for the interest to travel to.
If you ship anything internationally, this is the cheap lesson: pin your romanisation in the first line of every document you release. Do not leave it for the receiving side to guess.
Half of it wasn't written by a person
Then I hit this, at the bottom of one article:
本文为AI基于第三方数据生成，仅供参考
"This article was generated by AI from third-party data. For reference only."
A financial outlet, disclosing it in its own footer. A large portal then republished that AI-written piece — disclosure and all.
Another began:
很多朋友不知道【…】，今天小绿就为大家解答一下
"Many friends won't know about this — today Xiaolu will explain it for you."
That is content-farm boilerplate. The same site filed an AI-model story under its Automotive section, which no human editor does.
The real breakdown:
| Type | Count |
|---|---|
| Human-reported original | 1 (CNMO) |
| Major platforms carrying it | 2 (Tencent News, ZAKER) |
| Hong Kong outlet, own edit | 1 (TechRitual) |
| Self-declared AI-generated | 2 |
"Covered by six Chinese outlets" is technically true. One act of reporting fanning out six ways is more honest.
An AI model, covered by an AI, filed by an AI
This is the part I keep turning over.
An article about a model that places five attention mechanisms on a Latin square was read by a machine, summarised by a machine, republished by a machine. No human touched the middle of that pipeline.
And the machine got every number right.
It also got the company name wrong.
Machines copy what's in the table precisely. They don't verify what isn't in the table. "Is it really Bidraft?" requires a search, and the summariser didn't run one.
The first reader of your technical documentation is increasingly not a person. That should change how you write it:
- Numbers in tables, not prose
- Names pinned at the top, unambiguously
- Minimal room for interpretation
Writing precisely instead of impressively used to be an ethics question. It's now also a distribution question.
What I took away
Two feelings at once.
Our material crossed a border and a language without being distorted. That is genuinely good, and it happened because the source documents were boring and exact.
And half of that crossing was done by machines, while our name calcified wrong along the way.
Building the technology and getting the technology called by its right name turn out to be different jobs. The first one happens in the lab. The second one doesn't.
The model: FINAL-Bench/Aether-7B-5Attn — weights, data recipe, training code, full logs, checkpoints. Apache-2.0.
Original coverage: CNMO 手机中国 · Tencent News
Who built it: VIDRAFT — with a V — https://vidraft.net
Top comments (0)